EP-DConvFormer: an entropy-preserving and frequency-aware transformer for radar-based human activity recognition

This paper presents EP-DConvFormer, a radar-oriented transformer architecture designed to improve the efficiency and recognition performance of radar-based human activity recognition. Built upon a ConvFormer-style hybrid backbone, the model integrates an orthogonal Haar-based downsampling strategy, deformable spatial alignment, and frequency-adaptive sparse attention to retain informative radar cues while reducing redundant computation. We further examine the proposed design using entropy-based empirical analyses to characterize feature variation under downsampling and sparsification. Experiments on a self-collected ultra-wideband (UWB) dataset and a public FMCW dataset show that EP-DConvFormer achieves 99.1% accuracy under the current de-identified random-split protocol on the UWB benchmark and consistently outperforms competing methods in recognition performance and computational efficiency. Compared with the ConvFormer baseline, it improves accuracy by 6.2 percentage points on the UWB dataset. Compared with the strongest competing method on the UWB benchmark, EP-DConvFormer also achieves higher accuracy and recall.